Cooperative research on new speech recognition methods including hidden Markov models and neural networks

点击翻译按钮获取中文摘要

Primary outcomes are in the following :(a) in the field of hidden Markov modelSpeaker adaptation of continuous HMM, combination of HMM and segmental statistics, consideration on semi-continuous HMM, new matrix-based calculation method for HMM, automatic construction of context-dependent HMM, learing of HMM by generalized descendent method,(b) in the field of neural networkspoken word recognition by sequential neural network (neural Markov model), consideration on feed-forward neural network for pattern recognition (dimensionality reduction, estimation of probability density function). generalized sequential machine. theoretical analysis for approximation of continuous function by recurrent neural network(c) in the field of acoustic/phonology and feature extractionnew acoustic feature model and feature hierarchies, extraction of distinctive feature by neural network, evaluation of smoothed group delay spectrum distance measure.(d) in the field of language modelmodeling of natural language by bigram/trigram/HMM/stochastic CFG, continuous stochastic CFG, analysis of phenomena in spoken dialog, sentence generation for QA system.(e) in the field of continuous speech recognition systemcontext-free grammar driven time synchronous continuous speech recognition using HMM, segmented trellis HMM calculation algorithm for continuous speech recognition, LR - HMM based continuous speech recognizor